影像去模糊是電腦視覺領域長期關注的低階影像復原問題，在攝影、監控、自駕車感知與醫學影像等下游應用中具有重要實用價值。儘管深度學習近年大幅推進了去模糊網路的表現，但絕大多數近期提出的去模糊架構（如 NAFNet、MIMO-UNet、VmambaIR）幾乎只在運動模糊資料集（GoPro、HIDE、RealBlur）上評估，僅 Restormer 等少數方法報告了失焦模糊資料集（DPDD）上的結果。 本論文量化此差距，指出在相同訓練框架下，兩種模糊類型於峰值信號雜訊比（Peak Signal-to-Noise Ratio， PSNR）存在約 5.5 至 8 dB 的落差，顯示即使各自充分訓練，失焦模糊在現有架構下仍難修復。 本論文提出一個由兩個模組構成、對主幹網路更動極小的去模糊框架。模組一探索以空間遮罩引導去模糊。首先在香港中文大學的 Blur Detection 資料集上，訓練一個以 MiT-B2 為編碼器，並以通道空間注意力模組（Convolutional Block Attention Module， CBAM）強化解碼器的模糊區域預測網路，於該子任務上達到交並比（Intersection over Union， IoU）≈ 0.86、F-score ≈ 0.92。接著透過空間特徵轉換（Spatial Feature Transform， SFT）層將預測遮罩注入主幹網路，並比較三種 SFT 變體（Spectral、Multi-scale、Dual-Branch）。實驗發現，雖然遮罩預測本身準確，但其對去模糊主任務的增益有限，且在運動模糊上難以發揮作用；其中表現最佳者為頻率感知的 Spectral SFT，此一發現促使本論文轉向模組二。 模組二提出一個即插即用的頻域增強分支，由頻域增強模組（Spectral Enhancement Module， SEM）與空間-頻域協同注意力（Spatial-Frequency Cross-Attention， SFCA）兩個元件組成，安插於主幹網路解碼器最終特徵圖與輸出卷積之間，額外參數低於主幹網路的 0.25%，推論時僅增加一對快速傅立葉轉換（Fast Fourier Transform， FFT）與逆轉換（IFFT），不提高推論成本。 本論文在 NAFNet、MIMO-UNet、Restormer 與 VmambaIR 四個具代表性的架構上，於 GoPro 與 DPDD 資料集以同一基準評估，並以 PSNR 與感知相似度（Learned Perceptual Image Patch Similarity， LPIPS）為衡量標準。結果顯示，在 DPDD 所代表的失焦模糊上，加入頻域分支能對四個架構中的三個帶來感知品質提升：NAFNet 的 PSNR 提升 0.72 dB、LPIPS 改善 23%，VmambaIR 的 LPIPS 改善 29%。在 GoPro 所代表的運動模糊上則呈現與架構相依的結果：VmambaIR 的 PSNR 提升 0.42 dB，而 Restormer 因其既有的 channel attention 與本分支的 channel cross-attention 重疊，PSNR 下降 1.85 dB。此一規律（即分支在缺乏自身頻域／通道注意力推理的架構上助益最大、在已具備者上則互相干擾）是本論文最具資訊量的發現。最後，本論文與生成式模型 GPT-5.5 進行比較。生成式輸出雖具視覺說服力，但在 PSNR 上仍與本方法相差 13 至 17 dB，顯示在保真度關鍵的應用中，確定性復原方法仍有其必要性。;Image deblurring is a long-standing low-level restoration problem in computer vision， with practical value in applications such as photography， surveillance， autonomous driving， and medical imaging. Although deep learning has greatly improved deblurring networks in recent years， most recent architectures， such as NAFNet， MIMO-UNet， and VmambaIR， are tested almost only on motion-blur datasets (GoPro， HIDE， RealBlur)， and only a few methods such as Restormer report results on a defocus dataset (DPDD). This thesis quantifies this gap and shows that， under the same training setup， the two blur types differ by about 5.5 to 8 dB in Peak Signal-to-Noise Ratio (PSNR). Even when each is fully trained， defocus blur remains harder to restore. We propose a deblurring framework of two modules that change the backbone as little as possible. Module 1 explores spatial-mask guidance. A blur-region predictor is trained on the CUHK Blur Detection dataset， using MiT-B2 as the encoder and a Convolutional Block Attention Module (CBAM) to strengthen the decoder， reaching an Intersection over Union (IoU) of about 0.86 and an F-score of about 0.92. The predicted mask is then injected into the backbone through a Spatial Feature Transform (SFT) layer， and three SFT variants (Spectral， Multi-scale， and Dual-Branch) are compared. The mask is accurate， but its benefit to deblurring is small and it does not work on motion blur. The best variant is the frequency-aware Spectral SFT， which motivates Module 2. Module 2 is a plug-and-play frequency branch made of a Spectral Enhancement Module (SEM) and a Spatial-Frequency Cross-Attention (SFCA). It sits between the decoder′s final feature map and the output convolution， adds under 0.25% extra parameters， and uses only one Fast Fourier Transform (FFT) and one inverse transform (IFFT) at inference， so it does not raise inference cost. We evaluate the framework on four backbones， NAFNet， MIMO-UNet， Restormer， and VmambaIR， on GoPro and DPDD under one protocol， using PSNR and Learned Perceptual Image Patch Similarity (LPIPS). On defocus blur (DPDD)， the branch improves three of the four backbones: NAFNet gains 0.72 dB in PSNR and 23% in LPIPS， and VmambaIR gains 29% in LPIPS. On motion blur (GoPro)， the result depends on the architecture: VmambaIR gains 0.42 dB， while Restormer drops 1.85 dB because its own channel attention overlaps with the branch′s channel cross-attention. This rule， that the branch helps most when the backbone lacks its own frequency or channel-attention reasoning and interferes when it already has it， is the main finding of this thesis. Finally， we compare the framework with the generative model GPT-5.5. The generative outputs look convincing but still trail our method by 13 to 17 dB in PSNR， showing that deterministic methods remain necessary when fidelity matters.
